NVIDIA AI Podcast

As AI factories scale and token costs become a defining competitive variable, the way businesses measure infrastructure ROI needs to change. In this episode, Shruti Koparkar from NVIDIA's Accelerated Computing team breaks down tokenomics—the four-pillar framework of token utility, supply, demand, and monetization—and reveals why NVIDIA Blackwell's architecture delivers 50x more tokens per watt than NVIDIA Hopper, translating to a 35x reduction in token cost.

🔬Topics covered:

The four pillars of tokenomics: utility, supply, demand, and monetization

Why cost per token beats FLOPS per dollar as an infrastructure metric

NVIDIA Blackwell vs. Hopper: 50x more tokens per watt, 35x lower token cost

How extreme co-design turns spec-sheet numbers into real-world output

Jevons paradox: why lower token cost always drives more GPU demand, not less

The four business models for turning tokens into revenue

Chapters:

00:00 – Introduction and the four pillars of tokenomics

02:09 – Token value: intelligence, interactivity, and use case mapping

06:32 – Estimating token demand: users, reasoning, and agentic multipliers

10:00 – Token supply and why cost per token is the right infrastructure metric

13:12 – NVIDIA Blackwell vs. Hopper: 50x more tokens, 35x lower cost

14:52 – Extreme co-design for lowest token cost and the NVIDIA Vera Rubin platform

21:10 – How software multiplies hardware performance (8x gains in six months)

23:56 – Token monetization: pricing and business models

26:52 – Jevons paradox and the future of GPU demand

Snap processes more than 10 petabytes of experimentation data every single morning—and with NVIDIA GPU-accelerated Apache Spark on Google Cloud, Snap cut job costs by 76%, reduced memory usage by 80%, and eliminated 120 terabytes of disk spill from its pipelines.

Prudhvi Vatala, head of engineering platforms at Snap, joins the NVIDIA AI Podcast to break down how he and his team completely modernized data infrastructure for a social platform serving nearly a billion monthly active users—using NVIDIA cuDF plugin (formerly referred to as NVIDIA RAPIDS plugin) for Apache Spark on Google Kubernetes Engine, with zero application code changes.

🔬Topics covered:

How Snap runs A/B tests at planetary scale using rigorous statistical methods like heterogeneous treatment effect detection and variance reduction

Why Snap reuses idle inference GPUs between 1–5 a.m. for batch data processing—and how it built a Kubernetes-based platform to do it

How NVIDIA cuDF delivered 3x+ speedups on join-heavy Spark jobs with no code rewrites

The full business impact: 76% cost reduction, 62% fewer cores, 80% less memory, 120 TB of spill eliminated

How a three-way partnership between Snap, NVIDIA, and Google Cloud made it possible in just 8–9 months

Chapters:

0:00 Introduction and Snap overview

3:35 What is Snap’s experimentation platform?

4:05 Why experimentation, safety, and privacy are core at Snap

4:52 How A/B testing works at billion-user scale

8:14 Discovering NVIDIA cuDF plugin

9:06 Benchmarking results: join, union, and aggregation jobs

12:00 Reusing idle GPUs overnight via GKE

13:24 Building a bottom-up GPU data platform at Snap

17:48 Results: 76% cost reduction and partnership impact

20:56 Snap’s evolution and what’s next

Learn more:

NVIDIA cuDF: https://developer.nvidia.com/topics/ai/data-science/cuda-x-data-science-libraries/cudf#accel-apache

LangChain has surpassed 1 billion downloads—and the framework that started as a weekend project is now the harness powering the next generation of production-grade AI agents. In this episode, Harrison Chase, co-founder & CEO of LangChain, breaks down the architecture behind deep agents, explains why systems like Claude Code, Manus, and Deep Research all share the same foundational pattern, and lays out what it actually takes to deploy autonomous agents responsibly in the enterprise.

🔬Topics covered:

What is a "deep agent," and why does architecture matter more than ever?

How enterprises are (and aren't) embracing autonomous agents

LangSmith: observability, tracing, and evaluation-driven development

Mixing frontier and open models (NVIDIA Nemotron) in multi-agent systems

What's next: async subagents, proactive/always-on agents, agent memory, and agent identity

Chapters:

00:00 – LangChain origin story and the deep agent architecture

01:46 – What is a deep agent?

03:31 – Enterprise trust: risk, autonomy, and iteration

04:38 – LangSmith: observability and evaluation-driven development

13:30 – Frontier vs. open models and the Nemotron Coalition

18:10 – What's next: async subagents, agent memory, and agent identity

Generative AI can predict whether a plane takes off—but does it know why? Nicolas Cerisier, VP of 3DEXPERIENCE Platform R&D at Dassault Systèmes, explains how industrial world models go beyond pattern recognition to embed the actual laws of physics, chemistry, and engineering. In this episode of the NVIDIA AI Podcast, he also breaks down Dassault's three virtual companions (AURA, LEO, and MARIE), their 25-year collaboration with NVIDIA, and a stunning real-world use case: helping NIAR rebuild aircraft designs part by part, using AI.

What if one AI brain could run every robot on the planet—a humanoid, a warehouse arm, and a dog-like inspection bot—all at once?

That's not a thought experiment. That's what Skild AI is building right now.

Deepak Pathak (CEO and Co-Founder) and Abhinav Gupta (President and Co-Founder) of Skild AI join the pod to break down Skild Brain—a universal, general-purpose AI model designed to power robots of any form factor, tackling any task, from a single shared intelligence.